orgrctera/legalbenchrag_contractnli_qa

LegalBenchRAG: ContractNLI (Question Answering)

This dataset is the ContractNLI slice of LegalBench-RAG, packaged as a question answering (QA) task. Each example pairs a natural-language question about a specific non-disclosure agreement (NDA) with gold answer text: one or more verbatim snippets from the contract that support the answer. The formulation follows the same underlying corpus and LegalBench-style queries as the broader LegalBench-RAG benchmark, but here supervision is answer strings (contract passages), not retrieval-style character spans and file paths.

Background: ContractNLI

ContractNLI (ContractNLI: A Dataset for Document-level Natural Language Inference for Contracts, Koreeda & Manning, Findings of EMNLP 2021) introduced document-level natural language inference for contracts. Given a contract and a set of hypotheses (fixed statements about obligations, survival, confidentiality, notice, etc.), a system must decide whether each hypothesis is Entailed, Contradicted, or Not mentioned by the contract, and for non-neutral labels identify evidence spans (sentences or list items) in the document.

The original release covers 607 NDAs with 17 shared hypotheses annotated across documents. ContractNLI is widely used because of long, legally dense text, imbalanced labels, joint classification + evidence structure, and difficult phenomena such as negation by exception.

• Original resource: ContractNLI project page · GitHub: stanfordnlp/contract-nli
• License: CC BY 4.0 (see the official ContractNLI distribution for full terms).

Task: Question answering for ContractNLI

In this Hugging Face distribution, the task is QA: given input (a single question naming the agreement and asking about a clause-style obligation or definition—aligned with ContractNLI-style hypotheses), a model should produce the supporting contractual language that answers the question.

• expected_output is a JSON-encoded list of strings. Each string is a contiguous excerpt from the contract (often multiple fragments when the answer is spread across clauses or a heading plus a sub-clause). Evaluators typically compare model outputs to these references via exact match, token F1, or similar text-overlap metrics; retrieval-specific offsets are not required for this split.

Typical uses:

• Generative QA or extractive QA on legal NDAs with multi-span supervision.
• RAG evaluation where the generator must ground answers in contract text (this dataset supplies reference answers; retrieval of the right passage may be evaluated separately or jointly).
• Analysis of how well LMs handle fixed hypothesis templates over real NDAs.

This is not the original three-way NLI classification task in isolation; it is the QA packaging of ContractNLI-derived examples within the LegalBench-RAG family (exported from Langfuse; see Dataset structure).

Relation to LegalBench and LegalBench-RAG

• LegalBench is a broad suite of tasks for legal reasoning in language models; several tasks connect to ContractNLI-style settings.
• LegalBench-RAG (Pipitone & Alami, 2024) emphasizes evaluating retrieval in legal RAG with minimal relevant segments annotated by legal experts. This repository’s QA variant focuses on what text constitutes the answer as a list of strings, which pairs naturally with end-to-end QA or RAG systems that output cited language.

When citing work built on this dataset, cite ContractNLI for the underlying corpus and hypotheses lineage, LegalBench-RAG for the benchmark context, and this dataset card for the specific HF QA packaging.

Dataset structure

Field	Description
id	Unique example identifier (UUID string).
input	Natural-language question (typically names the NDA and asks a yes/no–style or definitional question consistent with ContractNLI themes).
expected_output	JSON string that parses to a list of strings—the gold answer passage(s) from the contract, in order.
metadata.item	JSON object mirroring export metadata (e.g. query and expected answer list; field names follow the upstream Langfuse export).

Splits: train and default each contain 977 rows (duplicate views of the same examples).

Example records

The following examples are taken from the published parquet; line breaks may be wrapped for display.

Example 1

• input:
  Consider Aspiegel's Non-Disclosure Agreement; Does the document allow the Receiving Party to acquire information similar to the Confidential Information from a third party?

• expected_output (parsed JSON list; two strings):

  [
    "The confidentiality undertakings at clauses 2 and 3 above shall not apply to any Confidential Information which the Recipient Party can prove:",
    "(c) was obtained legally from any third party, and is not the subject of any restriction as to its use or disclosure imposed by or on that third party at the time of provision; or"
  ]
  

Example 2

• input:
  Consider the Mutual Non-Disclosure Agreement between Wayne Fueling Systems and Unknown Party; Does the document state that Confidential Information shall only include technical information?

• expected_output (parsed JSON list; one string):

  [
    "Confidential Information also includes, but is not limited to, personal data as defined in this Agreement or by applicable law, whichever is broader, and personal data shall not be required to be marked “Confidential” or “Proprietary” to be treated as Confidential Information under this Agreement. "
  ]
  

References

1. ContractNLI (original dataset and NLI task)
   Yuta Koreeda and Christopher D. Manning. ContractNLI: A Dataset for Document-level Natural Language Inference for Contracts. Findings of EMNLP, 2021.
   ACL Anthology · PDF

   Abstract (short): The paper proposes document-level NLI for contracts: classify hypotheses as entailed, contradicted, or not mentioned, and identify evidence spans; it releases a large annotated contract corpus and strong span-based baselines.

   @inproceedings{koreeda-manning-2021-contractnli-dataset,
     title = "{C}ontract{NLI}: A Dataset for Document-level Natural Language Inference for Contracts",
     author = {Koreeda, Yuta and Manning, Christopher},
     booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2021},
     year = {2021},
     pages = {1907--1919},
     url = {https://aclanthology.org/2021.findings-emnlp.164}
   }
   

2. LegalBench-RAG (retrieval benchmark and RAG evaluation)
   Nicholas Pipitone and Ghita Houir Alami. LegalBench-RAG: A Benchmark for Retrieval-Augmented Generation in the Legal Domain. arXiv:2408.10343, 2024.
   arXiv · PDF

   Abstract (short): The paper argues for evaluating retrieval in legal RAG with minimal relevant segments annotated by legal experts, enabling precise metrics and better source attribution in LLM outputs.

3. LegalBench (umbrella benchmark)
   LegalBench — task collection and associated publications for broader legal NLP evaluation.

Community artifacts: zeroentropy-ai/legalbenchrag (reference tooling and materials related to LegalBench-RAG).

Acknowledgments

Data lineage traces to ContractNLI (Stanford NLP / Hitachi America, Ltd. distribution terms) and the LegalBench-RAG construction. Use of the original ContractNLI materials remains subject to CC BY 4.0 and any additional terms stated on the official ContractNLI release.